my seq2seq network for predicting phonetic features at from character strings after 10 epochs is at 99% accuracy on the validation set and pronounces (e.g.) "fediverse" (not in training set) almost flawlessly (I'll transcribe the features as "fidiverz") but seems to consistently mess up on interdental fricatives ("theorizing" comes out as "feruhzing," "lathe" comes out as "lat-tee," "this" comes out as "sis")
similar problems with /ʒ/ ("genre" comes out as ?ehnuh where <?> is a consonant described as a "voiced alveolar fricative stop" with a hint of velar thrown in). probably because these sounds combine be less than 1% of all sounds and might not be present more than a handful of times in the training set. I might have to think about partitioning differently or augmenting the data set to even out the distribution
@aparrish Have you considered that maybe your network is just German?
@sluglife it is amusing how easy it is to compare these mistakes to the mistakes of people who learn english as adults! (though I'm definitely not claiming that the processes are internally similar in any way)
@aparrish Yeah, it makes sense with how phonetically inconsistent English can be. Have you released the source somewhere? I'd love to try this myself and a friend of mine does computational linguistics and would probably also love to.
@sluglife not yet, still just playing around. hoping to make a demo and release code in a few weeks
@aparrish I tried to make that noise but it was a challenge.
@aparrish I don't know what these words are but they sound very impressive and I look forward to more good words
@wilbr yeah sorry for better or worse I live-toot my experiments on mastodon before it all goes through the filter of "try to make this accessible for people who are not allison"
@aparrish it's ok! It's better that way. Like hearing someone on Jeopardy say "what is trichonosis" and you're at home like "--chonosis! Yeah!!"
@aparrish this sounds super super cool and can't wait to see the demo :)
Server run by the main developers of the project It is not focused on any particular niche interest - everyone is welcome as long as you follow our code of conduct!